Reliable and Interpretable AI for Medical Imaging, Morteza Mardani, PhD

Abstract:
The advent of Artificial intelligence (AI) is arguably a renaissance for medicine. AI applications are fundamentally reshaping the healthcare landscape by optimizing the cost while improving patient outcomes. The widespread clinical adoption of AI for medical imaging, however, is hindered by challenges such as lack of data, lack of interpretability, and reliability concerns of deep learning algorithms.

The talk discusses these challenges broadly. In order to shed some light on the solutions, it then provides AI examples for fast MR imaging. Generative adversarial networks as an advocate for semi-supervised learning are discussed to deal with the data scarcity. To address reliability, variational autoencoders are presented to create uncertainty maps that could guide radiologists about the confidence of their interpretations. Finally, to provide interpretability, unrolled neural networks are presented that integrate side information into deep learning, and can be divided into subnetworks that are better explainable.

As long-term vision, the potential of integrating various clinical information sources from imaging and non-imaging data into deep learning to create robust decision support for diagnostic and therapeutic tasks is discussed.